Research map / Image generation / Fast diffusion sampling
Revisiting Feature Disentanglement Strategy in Diffusion Training and Breaking Conditional Independence Assumption in Sampling
Wonwoong Cho, Hareesh Ravi, Midhun Harikumar, Vinh Khuc, Krishna Kumar Singh, Jingwan Lu, David Iseri Inouye, Ajinkya Kale
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add joint guidance via GCDM (lambda=1.0) instead of independent CDM guidance (ablation) | FDiff · AFHQ | FID: 10.57 -> 8.58 | Improved |
| increase joint guidance weight lambda in GCDM (ablation) | FDiff · AFHQ | LPIPS: 0.59 (CDM) -> 0.57 (lambda=1.0) | Mixed |
| use GCDM instead of CDM without timestep scheduling (ablation) | FDiff · FFHQ | FID: 21.43 -> 14.46 | Improved |
| add timestep-dependent weight scheduling to CDM (ablation) | FDiff with CDM · FFHQ | FID: 21.43 -> 10.50 | Improved |
| add timestep-dependent weight scheduling to GCDM (ablation) | FDiff with GCDM · FFHQ | FID: 14.46 -> 10.21 | Improved |
| add timestep scheduling (ablation) | FDiff with CDM · FFHQ | LPIPS: 0.47 -> 0.31 | Hurt |
| use GCDM instead of CDM for Stable Diffusion text prompt composition | Stable Diffusion V2 · multi-object prompts (qualitative) | composition realism | Improved |
| lower lambda to weaken joint guidance (with beta_s=1) (ablation) | FDiff · FFHQ | FID: 11.99 -> 15.45 | Hurt |
| lower lambda to weaken joint guidance (with beta_s=1) (ablation) | FDiff · FFHQ | LPIPS: 0.34 -> 0.49 | Improved |
| use FDiff with two disentangled features instead of baselines for reference-based translation | FDiff · FFHQ | FID: 25.06 (SAE) -> 11.99 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DEADiff — DEADiff: An Efficient Stylization Diffusion Model with Disentangled Representations (CVPR 2024) · better on 1 of 1 shared result
- StarGAN v2 — StarGAN v2: Diverse Image Synthesis for Multiple Domains (CVPR 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Can Diffusion Models Disentangle? A Theoretical Perspective (NeurIPS 2025 · similarity 0.91)
- Isometric Representation Learning for Disentangled Latent Space of Diffusion Models (ICML 2024 · similarity 0.89)
- Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.89)
- Tuning Timestep-Distilled Diffusion Model Using Pairwise Sample Optimization (ICLR 2025 · similarity 0.88)
- Move Anything with Layered Scene Diffusion (CVPR 2024 · similarity 0.88)
- NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Models (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
